Prediction-based Redundant Data Elimination with Content Overhearing in Wireless Networks Haiying Shen † , Shenghua He*, Lei Yu ‡ and Ankur Sarker † † Department of Computer Science, University of Virginia * Department of Computer Science and Engineering, Washington University in St. Louis ‡ College of Computing, Georgia Institute of Technology
Outline • Introduction • System Design • Performance Evaluation • Conclusion 2
Outline • Introduction • System Design • Performance Evaluation • Conclusion 3
Introduction Internet Traffic How Netflix, YouTube, Hulu, and Amazon became the Internet. File sharing is another major source. 30% of the Internet is just a Copy of itself. Redundancy 4
Introduction Redundant data elimination Suppressing duplicated data transmission using Redundancy Elimination techniques: 1. Packet-based RE 2. Content-based RE Redundancy Elimination 5
Related work Wired network 1. WAN optimization [Riverbed Networks 2013] [Juniper Network 2014] 2. Server to client RE [Agarwal et al., NSDI 2010] [Hua et al., Infocom 2014] 3. Prediction-based RE [Zohar et al., Sigcomm 2011] [Yu et al., ICNP 2012] Wireless network 1. Packet-based RE [Hua et al., Infocom 2015] [Sanadhya et al., Mobicom 2012] 2. Content-based RE [Dogar et al., Mobicom 2008] [Afanasyev et al., NSDI 2008] 6
Introduction Problems in previous methods 1. Caches at sender and receiver would be outdated Disrupts RE’s correctness and degrade its performance 2. Overhearing probability estimation is difficult Consequently degrades the performance of RE Causes significant communication cost and complex coordination among nodes Solution: Prediction-based Redundancy Elimination 1. The receiver stores the received and overheard data stream in a chain of chunks. 2. It compares the chunks of the incoming packet with the stored chunk chains in the cache. 3. The receiver sends to the sender future data predictions that include the hashes of chunks on the chain. 7
Introduction Challenges of RE 1. Identifying duplicate chunks of hundreds of bytes at sub-packet level and work on lower-bandwidth wireless links. 2. Transmission cost of predictions has to be considered. 3. There are possibilities of missing data. Our method: PRECO 1. Redundancy elimination technique An effective, efficient and scalable solution for content- overhearing based IP-layer RE over wireless links. 2. Chunking and caching To divide the payload into evenly distributed chain of chunks. 3. Adaptive prediction algorithm To improve prediction accuracy and reduce the prediction overhead. 8
Outline • Introduction • System Design • Performance Evaluation • Conclusion 9
System Design Overview a b c d f a b c d f I c, H c AP1 AP2 AP1 AP2 I d, H d I e, H e data a a b c d e b I c I d f a c d e data C1 C2 C1 C2 a, H a c, H c d, H d e, H e a, H a c, H c d, H d e, H e Prediction and RE Overhearing and caching 10
System Design Chunking and caching The overall procedures divide into two steps: Chunking algorithm: PRECO divides the payload of a packet into several chunks. MAXP [1], a content-based chunking algorithm is used to define chunk boundary. Caching received and overheard chunks: Nodes overheard TCP streams and stored as chunks based on stream IDs (src, dst, src port, dst port). [1] Anand, Ashok, et al. "Redundancy in network traffic: findings and implications." ACM SIGMETRICS Performance Evaluation Review 37.1 (2009). 11
System Design Prediction-based RE Prediction of receive packets works as follow: PRECO determines one matching chunk as a prediction anchor. Chooses a chain of chunks for prediction based on highest matching length. Prediction chunk is chosen based on the matching degree. Virtual chunk is created based on matching degree. For a received packet, one chain of chunks is selected as prediction based on virtual chunk. Prediction transmission and shim decoding: A prediction windows is used to increase the efficiency. Receiver sends the chunk prediction in a prediction message. Upon receiving, the sender stores the prediction in cache. For an outgoing packet, the sender performs chunking using same algorithm and insert shim into the packet. Once receiving a packet containing shim, receiver finds the shim from sender. 12
System Design Adaptive prediction algorithm Size of the prediction window, W where P A is prediction chunk chain based on Anchor A R(P A ) is hit ratio of prediction PA W 0 is initial prediction window size N A is the next expected byte sequence number based on A N B is the next expected byte sequence number based on B d T prediction distance threshold 13
System Design Adaptive prediction algorithm Size of the prediction window, W When making predictions with a new prediction anchor: the algorithm first computes the hit ratio of previous prediction, and then accordingly adjusts the prediction window. 14
System Design Redundancy-aware source routing Gateway a b c e a b c d ETT Estimation: ETT=3 ETT=2 • Using “A1 -A3- A5”: 3+3+16 = 23 a b c d A2 A1 • Using “A2 -A4- A5”: 8+12+12 = 32 e ETT=4 ETT=3 e A4 A3 a b c d A5 ETT=3 ETT=4 a b c e C2 C1 a b c d 15
System Design Redundancy-aware source routing The redundancy-aware source routing: Routing metric: Redundancy Estimated Transmission Time 𝑡 1−𝛽 𝑆𝐹𝑈𝑈 = 𝐹𝑈𝑌 × 𝐶 where, ETX is Expected Transmission Count [1] S is average packet size B is the bandwidth α is average redundancy ratio Routing Protocol: Compute RETT matric for all links Apply Dijkstra’s shortest path algorithm to find path route with lowest RETT Without any overhearing consideration [1] De Couto, Douglas SJ, et al. "A high-throughput path metric for multi-hop wireless routing." Wireless Networks 11.4 (2005). 16
Outline • Introduction • System Design • Performance Evaluation • Conclusion 17
Experiment Simulation settings 1. Real trace collected using YouTube app over 3.5 GB data Two smartphones (Iphone 6 and Xiao Mi 3) connected with one laptop (Lenovo T420 Windows 10 machine) 2 different videos with similar contents, watched twice 60 minutes a day for 7 day Captured packets using Wireshark 2. 2 scenarios- One AP and one client without content overhearing Two Aps and two clients, with content overhearing Compared methods 1. EndRE: new finger print technique for end users 2. Asymmetric Caching (AC): RE operations based on feedback cache 3. REfactor: finer-granularity redundancy at the sub-packet level with content overhearing 18
Experiment RE efficiency Observation : RE efficiency follows PRECO>REfactor>EndRE>AC Reason : EndRE and AC do not support overhearing; REfactor uses overhearing probability estimation. 19
Experiment Content overhearing Observation : PRECO>REfactor>EndRE>AC Reason : The bandwidth saving is mainly caused by RE efficiency 20
Experiment Network overhead Observation : REfactor has the highest network overhead among all these RE methods Reason : The overhearing probability estimation results larger amount of network overhead 21
Experiment Simulation settings 1. we deployed a mesh network with 5 rows and 5 columns in total 1. First node is the gateway 2. Three clients associated in three distinct nodes 3. the overhearing coverage to 1 4. The average date rate for each link varies from 800Kps to 1200Kps 2. The gateway send traces to two different clients Compared methods 1. ETT-based routing: The gateway determines the optimal route to a receiver using ETT metric, and no network-wide PRECO deployment to perform RE. 2. Redundancy-aware routing without content overhearing: network-wide PRECO is deployed without content overhearing. 3. Redundancy-aware routing with content overhearing: network-wide PRECO with content overhearing. 22
Experiment Redundancy-aware routing Observation : Redundancy-aware routing with content overhearing produces more throughput Reason : The gateway steers the traffic through the nodes with high redundancy 23
Outline • Introduction • System Design • Performance Evaluation • Conclusion 24
Conclusions 1. we propose a prediction-based IP-layer RE method with content overhearing named PRECO for wireless networks. 2. We propose novel prediction algorithms that allow PRECO to effectively improve prediction accuracy and overall bandwidth saving. 3. Trace-driven simulation results show that PRECO provides significant performance benefits in comparison with other RE methods. Future work Further take into account efficiently learn the overhead data streams of all nodes for route determination in mesh networks. 25
Thank you! Questions & Comments? Ankur Sarker as4mz@Virginia.edu Ph.D. Candidate Pervasive Communication Laboratory University of Virginia 26
Recommend
More recommend
